At the AGI Playground in Beijing on June 21st, 360 Group Vice President and Head of NanoAI, Liang Zhihui, delivered a compelling presentation titled “The Next Super Entry Point in the AI Era: Super Search.”
Liang highlighted the persistent pain points inherent in traditional search. Users are often constrained by keywords, with the quality of search results deteriorating rapidly as keyword count increases, particularly beyond twenty. He noted that users rely on search engines primarily to find specific websites and resources, or to solve problems. The advent of large language models (LLMs) has drastically changed user behavior, shifting the focus from carefully crafted keyword searches to more natural language queries, with question-based searches now accounting for 90% of the total.
Recognizing the shift from keyword-centric search to intent-driven search, 360 has proactively leveraged AI to revamp its search capabilities. The company’s journey in AI search has been marked by three distinct phases.
The initial 1.0 version of AI search, mirroring strategies employed by other global search engines, utilized AI to pre-summarize answers to common queries, thereby placing the summary at the top of search results. However, this approach, fundamentally based on “memorization + retrieval”, was limited. It primarily relied on pre-defined summaries for frequently asked questions, which did not improve results as the “keywords” grew longer. This resulted in inaccurate answers, and a user experience that did not significantly improve on what was already available.
The 2.0 iteration saw the industry exploring deeper integration of LLMs with search, employing intent recognition to facilitate a more comprehensive, multi-intent-driven search to provide more complete information. While a more effective approach, the user experience remained similar to the previous generation.
Responding to decreasing model call costs and increasingly complex user queries, 360 has now advanced to the 3.0 phase of its AI search. This 3.0 search engine leverages a “task engine” with advanced autonomous task planning capabilities. When a user submits a request, the system can assess whether the need is simple or complex, employing a “step-by-step planning — step-by-step execution — step-by-step understanding” workflow to deliver results with the precision of a specialist.
Most recently, NanoAI was upgraded to a super search intelligent agent. When navigating complex problems, the search agent is capable of reasoning in a loop. By searching, assessing information, and verifying its output, the agent can understand and implement even the most sophisticated requests.
To illustrate, consider a user planning an outing for a group of friends with varying interests. The user may input a query to plan a rock-climbing trip for five friends of varying experience, asking for recommendations for premier rock-climbing locations with details on local cuisines and suggestions for nearby music festivals.
Such intricate requirements frequently stump traditional search engines that require a user to painstakingly compile keywords for “rock climbing,” “beginners,” “rock climbing and food,” and “rock climbing and music festivals” in an effort to find the right combination of search terms.
The NanoAI super search agent streamlines the process by eliminating the need for keyword manipulation. The user can succinctly articulate the need, no matter how complex, and NanoAI can utilize “intelligent questioning,” break down the complex meaning of the question and identify its user’s actual need.
The system then deconstructs the complex task into multiple subtasks, further breaking them down into multiple subtasks, then repeating the process as necessary.
As each subtask is fulfilled, NanoAI’s super search agent finalizes the request to deliver a visually rich and dynamic webpage with recommendations for rock-climbing sites, including a PDF and Word document format for the user to modify and adjust as needed. Each site’s recommendations include analyses of the climbing difficulty, suggestions for regional cuisines, music festival information, and other professional guide services with relevant information in anticipation of the user’s original requests.
According to Liang Zhihui, achieving these expert-level results involved three significant challenges:
* Effectively leveraging the capabilities of LLMs to decompose complex problems.
* Orchestrating multiple intelligent agents to collaboratively address complex tasks.
* Maintaining a high success rate in the face of extended execution chains.
These challenges were addressed by combining varied models. LLMs can be viewed as the brains of the operation, but in isolation, they can only execute simple tasks. To broaden the use cases, the NanoAI team built dedicated AI browsers, search and code generation tools, and intended to make it a completely domestic software model. Currently, the NanoAI super search agent has integrated multiple home-grown models. More than 80 models are working together, with models like deepseek r1 and Qwen 3 for advanced reasoning. Models with function call capabilities like Doubao and Qwen plus are used for function calls, while the team’s independently trained specialty models like those used for fast search and translation collectively create a critical intelligent cooperation system to achieve superior results.
For model deployment, NanoAI has solved the problem of model deployment to manage intricate documents. For example, it can analyze 500 very detailed documents to give answers, and this calls for the highest-level logical capability and performance. To ensure that its product and service functions correctly, the internal team has designed a suite of AI-based software.
Furthermore, both the mobile and PC versions of NanoAI offer intelligent agent creation. Users need only master prompt writing to invoke over 100 high-quality domestic and global MCPs built in to the platform. For certain paid MPCs, users have no need to register abroad. The platform supports integrated functionality. This design greatly lowers the barriers to access, and NanoAI’s mobile app enables users to build their custom agents based on their individual needs.
Liang Zhihui envisions that the NanoAI super search agent will propel the era of the super individual, enabling everyone to transcend their abilities and leverage AI to evolve from information recipients into value creators.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/3208.html
